Host-Model Extraction

Requires edge-studio 0.0.1rc22 or later.

The deterministic URL importer stores a page as one fact, or splits HTML table rows. For long prose pages where neither is enough, a stronger local Mac model can read the fetched material and propose discrete facts. The model only proposes — Edge validates every candidate against edge.demo.facts.v1 before anything is written to the store.

Run It

edge demo facts import-url "https://example.org/service-reference" \
  --store service_docs_v1 \
  --topic "service reference" \
  --tags service,docs \
  --extractor host-model \
  --extractor-model qwen3.5-27b-4bit \
  --json

• The extractor mode is explicit: without --extractor host-model, imports stay deterministic. There is no silent model fallback.
• --extractor-model names a local model; it defaults to qwen3.5-27b-4bit. If the model is not available locally, the command fails closed with a fetch remediation — it never reaches the network to find one.
• Host-model extraction works on single-URL imports only. crawl-url does not accept it.

What The Validation Layer Guarantees

Model output is treated as candidate facts only:

1. The model receives the fetched material plus a strict output contract.
2. Edge parses the response and normalizes only the top-level candidate wrapper.
3. Every candidate row passes the same deterministic edge.demo.facts.v1 validator used for file imports — required fields, types, and counts.
4. Invalid output fails closed before any store write. The model cannot redirect the target store.

Read The Receipt

The receipt marks this path as non-deterministic and fingerprints every stage:

Field	Meaning
extractor.mode	host-model
extractor.extractor_model_ref, model_path, model_sha256	Which local model ran, with a directory-manifest hash
prompt_sha256, schema_sha256	The exact instruction and output contract
input_sha256, model_input_sha256	The material given to the model
output_sha256, validated_payload_sha256	Raw model output and the validated payload
source_chars, model_input_chars, input_truncated	Whether the model saw the whole page: long pages are clipped at --max-chars, and the receipt says so truthfully
validation_status	passed — otherwise nothing was written
non_deterministic_extraction	true

Two different runs on the same page can produce different fact splits — that is the trade-off you opt into for extraction quality. If the receipt reports input_truncated: true, raise --max-chars or import the page in sections.

When To Use Which Extractor

Material	Use
Index page with an HTML table	Deterministic --split html-table-rows
Short page, one coherent topic	Deterministic single-page import
Long prose with many discrete facts	--extractor host-model
Several linked pages	crawl-url (deterministic only)